Vector-Valued Reward
Separates stochastic engagement from independent quality and moral-emotional outrage estimates[cite: 5].
Section 2 SpecTechnical Report / August 2026
Auditing
the Objective.
A Multi-Objective Bandit Framework and Reproducible Audit Protocol for Engagement-Based Recommendation Systems[cite: 1].
Exposing the mathematics behind algorithmic amplification and establishing concrete verification methods for independent auditors[cite: 1, 2].
Contextual Bandits
Algorithmic Amplification
Exposure Fairness
Reproducible Audit Protocols
01 / Research Context
Moving from diagnosis
to prescription.
How can platforms structurally bound extremity concentration without abandoning engagement?
This technical report extends "Engagement as Objective" into a robust engineering specification[cite: 1]. It outlines a contextual multi-armed bandit (CMAB) architecture featuring independent quality and moral-emotional outrage estimators[cite: 1, 5].
By connecting fair-ranking principles with automated audit checklists, the report bridges theoretical constraints and real-world compliance[cite: 1, 2, 8].
Read section 1 specification ↗02 / Core Architecture
Four pillars of the
audit framework.
Turning abstract fairness constraints into checkable protocol specifications[cite: 9].
Pareto-Optimal Filter
Implements a slate-level rejection rule bounding high-outrage exposure across rolling windows[cite: 1, 5, 10].
Section 4 FilterQuality-Aware TS
Combines a forced-exploration floor with Thompson Sampling to defeat degenerate feedback loops[cite: 1, 13].
Section 5 Bandits22-Item Checklist
Provides a five-domain framework for third-party auditors using public docs and logged data[cite: 2, 17, 21].
Section 6 Audit03 / Key Findings
Quantifying the
trade-off frontier.
Simulation performance across an 8-seed multi-run robustness evaluation[cite: 1, 2].
−21.1%
Outrage Exposure Reduction
Achieved via quality-aware Thompson Sampling over a 6,000-round simulated horizon[cite: 1].
+68.2%
Cumulative Engagement Regret
The explicit optimization price accepted to lower extremity concentration[cite: 1, 15].
Robustness Check
View simulation curves and charts ↗
Reported across independent random seeds with mean ± 1 SD bands to verify outcomes are structural rather than stochastic artifacts[cite: 2].
Citation & Access
Cite this work.
Use the BibTeX record below or download the full text PDF for complete equations, algorithms, and audit checklists.
Download PDF (26 Pages) ↓BibTeX Citation +
@techreport{satkynov2026auditing,
title = {Auditing the Objective: A Multi-Objective
Bandit Framework and Reproducible Audit
Protocol for Engagement-Based
Recommendation Systems},
author = {Satkynov, Ilgiz and
Abdyrasulova, Baian and
Krutchenko, Ivan},
year = {2026},
month = aug,
type = {Technical Report}
}